Moonshot AI Officially Launches Kimi K3, Ascend Day0 Support, U.S. Watches with Caution

Moonshot AI officially launched its latest open-weight model, Kimi K3, on Monday (27th), scaling the model’s parameter count to 2.8 trillion. This marks the world’s first open-weight model to enter the 3 trillion-parameter tier, immediately drawing widespread attention from the global tech community and international political arenas.

Technical Specifications and Performance

Kimi K3 adopts a Mixture-of-Experts (MoE) architecture with 896 experts, activating 16 experts (approximately 104 billion parameters) per token processed. The model features a native 1 million token context window and integrates a hybrid attention mechanism called Kimi Delta Attention (KDA), delivering higher computational efficiency in long-sequence processing.

According to Moonshot AI’s technical report, K3 achieves approximately 2.5 times higher intelligence efficiency per computation compared to its predecessor K2.

In evaluations by independent benchmarking firm Artificial Analysis, Kimi K3 ranks third globally in intelligence index, trailing only Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol. In coding performance, K3 secured first place on the blind testing platform Frontend Code Arena, demonstrating strong competitiveness in engineering implementation.

Commercial Licensing and Hardware Compatibility

Unlike traditional open-source licenses, Kimi K3 uses a custom “Kimi K3 License.” Moonshot AI mandates that commercial products or API providers with annual revenue exceeding $20 million must sign separate commercial agreements—a move seen as a strategy to prevent large cloud providers from free-riding.

On the hardware front, Huawei Ascend CANN announced compatibility on launch day, supporting inference and training on its Atlas A3 series hardware, enabling “domestic super-node” software-hardware co-optimization. Additionally, K3 supports deployment on NVIDIA GPUs and other general-purpose GPUs (GPGPUs).

Expert Opinions and Geopolitical Controversy

The release of Kimi K3 has sparked intense debate in Washington over AI regulation. Michael Kratsios, Director of the White House Office of Science and Technology Policy, publicly accused Moonshot AI of replicating U.S. AI advancements through “covert industrial-scale distillation” techniques.

Dean Ball, OpenAI’s Head of Strategy for the Future, described this open-weight approach as “AI communism,” yet acknowledged K3 as a “very good model” whose performance cannot be explained by distillation alone.

In contrast, a coalition of 25 tech giants including NVIDIA, Meta, and Microsoft jointly issued a statement opposing blanket restrictions on open-weight models. The statement argues that relying solely on closed-source models does not guarantee safety, and over-regulation could harm innovation.

Following the release and performance benchmarks of K3, market confidence in closed-source model pricing has wavered, with NVIDIA’s stock price dropping as much as 5%. Meanwhile, Moonshot AI’s valuation has surged from $20 billion to $50 billion, with reports suggesting preparations for an IPO in Hong Kong.